To study foraging decisions at high resolution, we filmed and quantified the behavior of individual animals on small lawns of bacterial food (~3 mm diameter) on an agar surface (Figure 1A and B). During a 40-min assay, animals spent ~97% of their time on the lawn, and preferentially occupied the edge of the bacterial lawn where bacterial density was highest (Figure 1C, Figure 1—figure supplement 1A–D). As a result, the tip of the animal’s head encountered the lawn boundary approximately once per minute. A few of these encounters resulted in a lawn leaving event, in which the animal exited the bacterial lawn (Figure 1D, Figure 1—video 1). Lawn-leaving occurred in 26% of the assays, with a mean value of one event per 95 min (Figure 1H–I, Figure 1—source data 1).

Once an animal left the lawn, it typically remained outside of the food for 1–2 min before re-entering the lawn (Figure 1—figure supplement 1E, Figure 1—source data 1). Most edge encounters were not followed by leaving; instead, animals that poked their head outside of the lawn either continued forward locomotion (head poke-forward), paused (head poke-pause), or executed a reversal (head poke-reversal; Figure 1E–G, Figure 1—video 1). Head poke-reversals were the most common event after lawn edge encounter (1.1 events/min), followed by head poke-forward and head poke-pause events; lawn leaving events were the least frequent (Figure 1I, Figure 1—source data 1).

High resolution analysis revealed characteristic behavioral dynamics prior to different events. Each head poke encounter with the lawn edge was preceded by an increase in speed over three seconds, and resolved by a reversal, pause, or continued forward movement within 5 s (Figure 1J). Lawn leaving events encompassed a longer acceleration that began 30 s before the leaving event (Figure 1J). At longer time scales, the average speed over 1 min before and after the edge encounter varied slightly for different classes of events, with head poke-forward events and lawn-leaving associated with the highest speeds. The different behavioral patterns preceding lawn leaving and head pokes suggest that lawn leaving is a distinct foraging decision, and not a random resolution of an edge encounter. In particular, it suggests that lawn leaving is linked to a persistent behavioral state reflected in ongoing locomotion speed.

To gain further insight into the behavioral states preceding lawn leaving, we coarse-grained our behavioral measurements into 10 s intervals (‘bins’) and expanded the time axis to allow analysis across different durations. On average, a slow rise in speed began two minutes before lawn leaving, with a rapid acceleration in the last minute before leaving (Figure 2A). We analyzed these behaviors in the context of the well-characterized roaming and dwelling states (Ben Arous et al., 2009; Flavell et al., 2013; Fujiwara et al., 2002). Training a two-state Hidden Markov Model (HMM) based on absolute speed and angular speed (reorientations) recovered a bifurcation of roaming and dwelling states (Figure 2—figure supplement 1A–D; Methods). Using these criteria, wild type animals on small lawns roamed for a median of 12% of the duration of the assay (Figure 2—figure supplement 1C, Figure 2—source data 1). Although 80% of lawn leaving events occurred within roaming states, less than 1% of the 10 s roaming bins contained a lawn leaving event (Figure 2B). Five minutes before leaving, 34% of animals were roaming, a fraction that steadily increased such that 80% of animals were roaming immediately before leaving (Figure 2C, Figure 2—source data 1). By contrast, the fraction of animals roaming before head poke-reversals increased only slightly before the event (Figure 2—figure supplement 3C). Within roaming states, the average speed was constant until the rapid acceleration ~30 s prior to lawn leaving (Figure 2D). To ask if the brief acceleration accounts for the apparent increase in roaming before lawn leaving, we compared the duration of roaming states that preceded lawn leaving to those that did not. We found that roaming states directly before lawn leaving were slightly longer than other roaming state, arguing that acceleration events alone do not drive the correlation between roaming and lawn leaving (Figure 2—figure supplement 1E).

The roaming-dwelling HMM was initially developed on uniform bacterial lawns (Flavell et al., 2013), which elicit more roaming than the small lawns used here (Figure 2—figure supplement 1F–H, Figure 2—source data 1). Expanding this model to include posture information identified eight subclasses of dwelling states (Cermak et al., 2020). To explore alternative analysis methods, we generated an HMM behavioral state model trained on our experimental conditions. An Autoregressive Hidden Markov model (AR-HMM) (Buchanan et al., 2017; Linderman et al., 2020), trained over an expanded set of locomotion parameters, converged on four states that largely reflect forward locomotion speed and related features (Figure 2—figure supplement 2A–K, Methods). These four states provided a different segmentation of locomotory arousal that partly overlaps with roaming and dwelling states (Figure 2—figure supplement 2M–N). Most lawn leaving events occurred during the high speed, high arousal AR-HMM state 3 (Figure 2E). Animals spent a median of 9% of the time in state 3 (Figure 2—figure supplement 2K, Figure 2—source data 1), but 5 min before leaving, 26% of animals were in state 3, which ramped to 82% immediately before leaving (Figure 2F, Figure 2—source data 1). Like roaming, state 3 increased only slightly before head poke reversals (Figure 2—figure supplement 3E). Animals across all behavioral states – roaming, dwelling, and the four AR-HMM states – were similarly distributed across the bacterial lawn, with strong enrichment near the lawn boundary (Figure 2—figure supplement 4A–D).

Both the two-state model and the four-state model show that arousal states increase before lawn leaving over at least two phases: an enrichment of a high arousal state 3–5 min before leaving, followed by a rapid acceleration in the last 30 s. Because the widely used two-state roaming-dwelling model captured the key features of lawn-leaving as well as these alternative methods, we applied that analysis to subsequent experiments.

Animals roam more and leave lawns more frequently when bacteria are hard to ingest or have low nutritional value (Ben Arous et al., 2009; Shtonda and Avery, 2006). Conversely, simply increasing food concentration suppressed roaming and leaving on small lawns (Figure 3A–B, Figure 3—source data 1; Bendesky et al., 2011). To separate the behavioral effects deriving from sensation of food odors from ingestion of food, we treated E. coli OP50 bacteria with aztreonam, an antibiotic that renders bacteria inedible to C. elegans by preventing bacterial cell septation (Ben Arous et al., 2009; Gruninger et al., 2008). Both lawn leaving and roaming were dramatically increased on aztreonam-treated E. coli lawns (Figure 3C–D, Figure 3—figure supplement 1A–B, Figure 3—source data 1, Supplementary file 2). Animals on aztreonam-treated lawns maintained the characteristic speed acceleration in the last 30 s before leaving (Figure 3—figure supplement 1C).

Locomotion speed is regulated by direct interoception of pharyngeal pumping, the motor behavior associated with intake of bacterial food (Rhoades et al., 2019). To ascertain whether similar mechanisms affect arousal and lawn leaving, we acutely inhibited pharyngeal pumping by optogenetic activation of the red-shifted channelrhodopsin ReaChR in pharyngeal muscle (Lin et al., 2013; Figure 3E–J). Acute feeding inhibition strongly stimulated both roaming and lawn leaving (Figure 3E–H, Figure 3—figure supplement 1D–E, Figure 3—source data 1, Supplementary file 2).

While roaming and head pokes began immediately upon feeding inhibition, leaving events did not occur immediately but instead accumulated throughout the 10-min stimulation interval (Figure 3I). In each case, leaving was preceded by a 30 s acceleration in speed (Figure 3J). Thus, feeding inhibition elicits both roaming and lawn-leaving behaviors, and lawn-leaving occurs probabilistically during roaming states.

To further examine the relationship between arousal states and lawn leaving, we examined mutants with known alterations in roaming and dwelling. Animals deficient in serotonin (tph-1), dopamine (cat-2), or the neuropeptide receptor NPR-1 (npr-1) roam at a higher rate than wild type (Cermak et al., 2020; Cheung et al., 2005; Flavell et al., 2013; Omura et al., 2012; Sawin et al., 2000). We found that each of these mutants showed increased lawn leaving compared to wild-type controls, strengthening the observed correlation between roaming and leaving rates (Figure 4A–C and E–G, Supplementary file 2). In addition, the fraction of animals roaming increased over several minutes before leaving events, as was observed in wild type animals (Figure 4I–K). Finally, each mutant accelerated during the 30 s prior to leaving (Figure 4M–O). Thus, although the molecular basis of arousal is different in each of these mutants, the overall dynamics of roaming and lawn leaving are preserved across genotypes.

Animals lacking the G-protein-coupled receptor Pigment Dispersing Factor Receptor (PDFR-1) roam less than wild type (Flavell et al., 2013; Ji et al., 2021; Meelkop et al., 2012). On the small lawns used here, pdfr-1 mutants roamed less, but left food lawns at the same low rate as wild type (Figure 4D and H). Roaming rates increased similarly during the 3 min prior to lawn leaving in wild type and pdfr-1 animals, suggesting that the coupling of roaming and leaving does not require PDFR-1 signaling (Figure 4L). Although their basal locomotion speed is lower (Flavell et al., 2013; Ji et al., 2021), pdfr-1 did accelerate slightly before lawn leaving (Figure 4P). In summary, neuromodulatory mutants varied in the fraction of time spent roaming and dwelling, but in each case lawn-leaving behaviors were coupled to roaming and a brief speed acceleration.

Optogenetic inhibition of feeding elicited immediate roaming and probabilistic lawn leaving in both pdfr-1 and tph-1 mutants (Figure 4—figure supplement 1), indicating that neither of these neuromodulators is essential for interpreting feeding inhibition. This was surprising because the tph-1-expressing NSM neurons sense bacterial ingestion and signal food availability via serotonin release (Rhoades et al., 2019), and were therefore candidates to relay feeding signals. NSM may act through additional transmitters as well as serotonin, or it may be redundant with additional neurons that detect feeding inhibition.

As a complement to the neuromodulatory mutants, we employed a circuit-based approach to manipulate arousal levels and examine effects on lawn leaving. Roaming is strongly stimulated by pdfr-1; we defined sites of pdfr-1 expression that stimulate roaming using an intersectional Cre-Lox system that restores pdfr-1 expression in targeted groups of cells in pdfr-1 mutant animals (Figure 5—figure supplement 1, Figure 5—figure supplement 2A; Flavell et al., 2013). Previous work identified a moderate effect of AIY, RIM, and RIA neurons as mediators of pdfr-1-dependent roaming (Flavell et al., 2013). We observed a stronger rescue of roaming upon pdfr-1 expression solely in the RIB neurons, which are active during rapid forward locomotion (Figure 5—figure supplement 2B–D; Ji et al., 2021; Wang et al., 2020).

Following this result, we asked whether optogenetic activation of RIB might be sufficient for roaming in wild-type animals. PDFR-1 signals through the heterotrimeric G protein Gαs to increase cAMP levels (Janssen et al., 2008), and optogenetic activation of groups of PDFR-1-expressing neurons with the bacterial light-activated adenylyl cyclase BlaC increases roaming (Flavell et al., 2013; Ryu et al., 2010). We found that acute optogenetic activation of only RIB with BlaC induced immediate roaming in over 80% of animals (Figure 5A–C, Figure 5—figure supplement 3A–B).

RIB::BlaC stimulation also potentiated lawn leaving, with over 70% of animals leaving lawns during the stimulation period (Figure 5D–E). Like lawn-leaving during optogenetic feeding inhibition, this behavior was probabilistic across the 10-min light pulse (Figure 5F). Animals accelerated slightly in the 30 s before leaving, but the speed increase appeared less pronounced than in other manipulations, suggesting that RIB activity may partially occlude the acceleration motif in lawn leaving (Figure 5—figure supplement 3F). Neither roaming nor lawn leaving was potentiated by blue light exposure in wild type control animals not expressing BlaC (Figure 5C, Figure 5—figure supplement 3A, C-D).

In addition to the neuromodulatory arousal systems, multiple food- and pheromone-sensing chemosensory neurons affect roaming, dwelling, and leaving behaviors (Supplementary file 1). Most of these neurons use the cyclic nucleotide-gated channel gene tax-4 for sensory transduction, and tax-4 mutants have diminished roaming behavior (Ben Arous et al., 2009; Fujiwara et al., 2002; Figure 6A, Supplementary file 1). However, we found that tax-4 mutant animals continued to leave lawns – indeed, they left at slightly higher rates than wild-type animals (Figure 6B, Supplementary file 2). Moreover, the temporal relationship between roaming and leaving was altered in tax-4 mutants, which typically roamed for only ~1 min prior to lawn leaving (Figure 6C, Figure 6—figure supplement 1A–B). These results indicate that loss of tax-4 disrupted the characteristic arousal dynamics associated with leaving.

Many of the 15 classes of sensory neurons that express tax-4 have been implicated in roaming or leaving behaviors (Supplementary file 1). We rescued tax-4 separately in AWC, which senses food odors; ASK, which senses amino acids and pheromones; ASJ and ASI, which sense pheromones, food, and toxins; and URX/AQR/PQR, which sense environmental oxygen (Ben Arous et al., 2009; Bendesky et al., 2011; Greene et al., 2016; Milward et al., 2011). Significant effects on roaming or leaving behavior were observed upon ASJ, ASK, or AWC rescue (Figure 6A–C, Figure 6—figure supplement 1 and S2). The strongest effects resulted from tax-4 rescue in the ASJ neurons, which partially restored roaming levels before lawn leaving (Figure 6A–C, Figure 6—figure supplements 1 and 2), and tax-4 rescue in the ASK neurons, which paradoxically suppressed leaving to a level below that of either wild type or tax-4 animals (Figure 6—figure supplement 2). Because ASJ and ASK showed opposite effects in these experiments, we also examined strains in which both neurons were rescued. Combined ASJ and ASK rescue normalized roaming and leaving compared to ASK rescue alone, albeit not to fully wild-type levels (Figure 6—figure supplement 2). While we have not tested all neurons and combinations, these results suggest that multiple tax-4-expressing sensory neurons have roles in arousal-related behaviors and highlight ASJ as a regulator of roaming and leaving dynamics.

Next, we asked how tax-4 mutants responded to acute inhibition of feeding. As in wild-type animals, optogenetic inhibition of pharyngeal pumping resulted in an immediate and strong increase in roaming in tax-4 mutants (Figure 6D, Figure 6—figure supplement 3A). However, feeding inhibition in tax-4 mutants increased leaving only slightly, unlike in the wild-type (Figure 6E, Supplementary file 2). Both of these effects were rescued by expressing tax-4 in ASJ neurons (Figure 6C–E, Figure 6—figure supplement 3B–D). Similarly, tax-4 animals on inedible food roamed at the same high rate as wild type animals but produced significantly fewer lawn leaving events (Figure 6—figure supplement 4); rescuing tax-4 in ASJ restored lawn leaving to wild-type levels. Thus tax-4 sensory mutants uncouple leaving behavior from its normal context in multiple ways: they can leave edible food lawns without an extended roaming state, and they are less likely to leave when feeding is inhibited, even while they are roaming.

From an ethological perspective, lawn-leaving is a classical foraging decision based on an animal’s assessment of the quality of a food patch (Charnov, 1976). Previous studies of lawn-leaving have identified features of the environment, like food availability, population density, and toxic repellents, that affect leaving probability as predicted by foraging theory. Here, we have extended these observations using the framework of computational neuroethology (Datta et al., 2019), examining behavior in detail across time to identify rare leaving events and the behavioral context in which they occurred. We found that leaving was tightly coupled to roaming both during spontaneous behaviors and after acute manipulations that induced roaming states. In each case, and in neuromodulatory mutants that altered roaming frequency, leaving occurred when animals were roaming, occurred probabilistically over minutes during the roaming state, and was preceded by a brief 30 s acceleration in speed. The stereotyped features of leaving behavior, which were not shared by other lawn edge encounters, suggest that it is a discrete behavioral motif associated with roaming states. Leaving can be viewed as a form of decision-making—a discrete action that drives an adaptive behavioral choice between alternatives, influenced by internal states such as arousal (Kennedy et al., 2014).

Roaming and dwelling comprise a widely used framework for defining arousal states in C. elegans, although changing experimental conditions or analysis methods can reveal sub-states or alternative classification frameworks (this work and Cermak et al., 2020; Gallagher et al., 2013; Raizen et al., 2008; Van Buskirk and Sternberg, 2007). Classical ethology presents such internal states as hierarchical and mutually exclusive (Tinbergen, 1951), while molecular and circuit analysis has extended and refined these ideas (Asahina et al., 2014; Flavell et al., 2013; Hindmarsh Sten et al., 2021; Hong et al., 2014; Moy et al., 2015; Pan et al., 2012). Our results indicate that lawn leaving behaviors are coupled to the roaming state, but within this state they are relatively rare and apparently probabilistic. More detailed studies may reveal additional regulators, such as an individual’s sensory experience, that shape the pattern of leaving behavior (Coen et al., 2014).

Our results identify two distinct biological mechanisms that regulate lawn leaving. First, leaving is coupled to arousal state. Leaving rates are 20-fold higher in roaming animals than in dwelling animals, and the leaving rates of most arousal mutants are largely explained by the fact that they spend more time roaming (Supplementary file 2). The second mechanism is controlled by tax-4 sensory neurons, which shape the behavioral dynamics of roaming and leaving across a range of conditions and stimulate leaving during feeding inhibition.

At a straightforward level, sensory neurons are well-placed to evaluate environmental quality in the framework of foraging theory. For example, ASK and AWC neurons sense amino acids and food odors (Bargmann, 2006), while ASJ neurons sense pheromones and toxins (Greene et al., 2016; Meisel et al., 2014), and other sensory neurons detect distinct environmental features (Bargmann, 2006). Accordingly, many sensory neurons affect roaming, dwelling, or leaving (Supplementary file 1), and their relative importance varies with context, such as the presence of chemical repellents or pheromones that signal population density (Dal Bello et al., 2021; Greene et al., 2016; Milward et al., 2011; Pradel et al., 2007). Under the conditions used here, ASK neurons suppressed and ASJ neurons promoted roaming and leaving, respectively. Rescuing tax-4 in ASJ partially restored features of normal leaving dynamics, including roaming before leaving.

Others have shown that chronic feeding deprivation across hours drives lawn leaving through the action of tax-4 in several sensory neurons (Olofsson, 2014). An unexpected result obtained here was that tax-4 sensory neurons were required to drive the high leaving rates after acute inhibition of feeding. Together, these results suggest that the sensory neurons are sites at which internal feeding cues, behavioral states, and specific foraging decisions are integrated. Additional experiments are needed to define the full set of sensory neurons that couple feeding inhibition to leaving, the nature of the interoceptive signal from pharyngeal pumping, and the readout of the sensory neurons. Pharyngeal interoception is mediated in part by the serotonergic NSM neurons (Rhoades et al., 2019), but we found that serotonin was not essential for the effects of acute feeding inhibition. With respect to interoception, some of the tax-4-expressing sensory neurons detect tyramine, one of several biogenic amines that regulate lawn leaving (Bendesky et al., 2011; O’Donnell et al., 2020), and many detect neuropeptides as well (Hapiak et al., 2013).

Sensory neurons in C. elegans and other animals are most often studied in the context of rapid behavioral responses, but they also have critical roles in endocrine signaling. For example, intrinsically photosensitive retinal ganglion cells drive circadian entrainment in mammals (Hattar et al., 2002) and in cichlid fish, visual stimuli from females are sufficient to induce an endocrine androgen response in males (O’Connell et al., 2013). We speculate that the combination of endocrine signals, neuropeptides, and fast transmitters released by sensory neurons couples roaming states to leaving states. The insulin peptides and DAF-7-TGF-beta protein that regulate roaming, dwelling, and leaving are primarily produced by tax-4-expressing sensory neurons, so these neurons bridge sensory and endocrine regulation of foraging (Taylor et al., 2021). Expression of these genes is regulated by pheromones and metabolic conditions, providing an additional layer beyond neural activity for long-term regulation of behavioral state. Many tax-4 sensory neurons also release classical neurotransmitters such as glutamate (ASK, AWC) and acetylcholine (ASJ), and therefore have the potential to modulate rapid behaviors such as the lawn-leaving motif (Taylor et al., 2021). Future studies can test this hypothesis while determining the mechanisms by which sensory neurons integrate signals to trigger lawn leaving.

Bacterial food used in all experiments was E. coli strain OP50. Nematodes were grown at 20 °C on nematode growth media plates (NGM; 51.3 mM NaCl, 1.7% agar, 0.25% peptone, 1 mM CaCl2, 12.9 μM cholesterol, 1 mM MgSO₄, 25 mM KPO4, pH 6) seeded with 200 μL of a saturated E. coli liquid culture that had been grown at room temperature for 48 hr or overnight at 37 °C (without shaking) from a single colony of OP50 in 100 mL of sterile LB (Brenner, 1974). All experiments were performed on young adult hermaphrodites, picked as L4 larvae the evening before an experiment. Wild-type controls were the CGC1 (previously PD1074) sequenced strain derived from the N2 Bristol strain (Yoshimura et al., 2019), except for Figure 4 tph-1 and cat-2 controls, which were the CX0001 isolate of the N2 Bristol strain. Mutant strains were backcrossed into wild type to reduce background mutations. Transgenic strains were always compared to matched controls tested in parallel on the same days. Full genotypes and detailed descriptions of all strains and transgenes appear in Supplementary file 3: Strain details.

Strains tested for pdfr-1 rescue using PCR-amplified genomic fragments (Figure 5—figure supplement 1) were from Flavell et al., 2013. For cell-selective pdfr-1 rescue, an inverted cDNA under the pdfr-1 distal promoter was the floxed rescue construct (Figure 5—figure supplement 2), and Cre expression was driven by a pan-neuronal promoter (tag-168), in RIM (tdc-1), in AIY and others (mod-1), in RIA (glr-3), and in RIB (sto-3). A 972 bp region upstream of the sto-3 gene that drives expression solely in the RIB neurons was cloned using these primers:

Strains and extrachromosomal arrays for tax-4 rescue were reported in Macosko et al., 2009 and Worthy et al., 2018. The sequences of the promoter ends are given below, along with the concentration at which the plasmid was injected for generating transgenic lines.

Transgenic animals were generated by microinjection of DNA containing the genetic construct of interest, a fluorescent co-injection marker (myo-2p::mCherry, myo-3p::mCherry, elt-2p::nGFP, elt-2p::mCherry), and empty pSM vector to reach a final DNA concentration of 100 ng/μL. Transgenes were maintained as extrachromosomal arrays.

For all assays, E. coli OP50 was grown overnight in a shaking LB liquid culture from a single colony at 37 °C. On the morning of the assay, 400 μL of saturated liquid culture was diluted into 5 mL of LB and allowed to grow to OD1 at 37 °C (~1.5 hours), as measured by spectrophotometer. The liquid culture was then spun down and resuspended in M9 buffer (3 g KH₂PO₄, 6 g Na₂HPO₄, 5 g NaCl, 1 ml 1 M MgSO₄, H₂O to 1 liter) then concentrated to a density of OD2 (and OD1 or OD4 in Figure 3A–B). To generate the test lawns, 2 μL of this concentrated bacterial resuspension was seeded onto NGM agar in the center of each well of a custom-made laser-cut six-well plate, where each well is 10 mm in diameter (Stern et al., 2017).

50 μL of bacterial resuspension was seeded onto a separate NGM agar plate to be used as a food density acclimation plate. Lawns were grown at 20–22°C for 2 hr before the assay. Adult hermaphrodites picked as L4s 16–20 hr before the assay were then transferred to acclimation plates. After 45–90 min, animals were transferred to an unseeded NGM plate, cleaned of E. coli, and transferred singly into each well of the assay plates, where they were placed on bacteria-free agar and allowed to find the small food lawn on their own. Animals of the same genotype were grouped on the same six-well plates and each plate was recorded by a single camera. We used 12 cameras, enabling simultaneous recording of up to 72 individual animals at a time. Temperature and relative humidity within the behavioral recording apparatus were continuously monitored during recordings to ensure that environmental conditions were consistent across filming locations. As a further precaution, the filming locations of each genotype and wild type controls within the recording apparatus were randomized across batches of experiments and days to average out behavioral influences deriving from non-uniform local environmental conditions. Assays were recorded for 1 hr at 3 frames per second using 12 8.8 MP USB3 cameras (Pointgrey, Flea3) and 35 mm high-resolution objectives (Edmund Optics). LED backlights (Metaphase Technologies) provided uniform illumination of the assay plates. Commercial software (Flycapture, Pointgrey) was used to record the movies.

Assays testing worm behavior on uniform bacterial lawns were performed as in the small lawn foraging assay with the exception that instead of 2 μL of bacteria seeded in the center of the well, 15 μL of OD2 bacterial suspension was spread evenly throughout the well. Bacteria were grown for 2 hr before acclimation and starting the assay.

Pharyngeal pumping was inhibited by expressing the red-shifted channelrhodopsin ReaChR (Lin et al., 2013) under the myo-2 pharyngeal muscle promoter (a strain generously provided by Steve Flavell). Animals were stimulated while navigating small food lawns described above. Experimental animals were grown on bacterial lawns containing 50 μM all-trans retinal (+ATR) overnight before assays. Control animals were placed on lawns made in parallel that did not contain retinal (-ATR). A 590 nm Precision LED with Uniform Illumination (Mightex) controlled with custom MATLAB software was used to deliver optogenetic stimuli. Animals were acclimated to small lawns (no ATR) for 20 min before being exposed to alternating 10 min intervals of light OFF and light ON using 590 nm light at 60 μW/mm² strobed at 10 Hz with a 50% duty cycle (Figure 3, Figure 4—figure supplement 1, Figure 6, Figure 6—figure supplement 3). Recording hardware and software was identical to that of off-food foraging assays without optogenetic stimulation except that 475 nm short-pass filters were used on recording optics (Edmund Optics) to prevent overexposure of the video recording during light pulse delivery.

For behavioral quantification in Figure 3E–F, Figure 4—figure supplement 1A and C–D, and Figure 6—figure supplement 3A, only data before, during and after the first light pulse is shown. For aggregate comparisons of within-animal behavior in light OFF vs. light ON conditions, the two light OFF and light ON pulses were merged, that is merging intervals 0–10 min +20–30 min, and 10–20 min + 30–40 min for statistical analyses. For statistical tests comparing the fraction of animals roaming across genotypes or conditions, steady state light intervals were used for averaging and comparing as indicated in the figure legends: 0–10, 12–20, 22–30 min.

Activation of the RIB neurons was accomplished using blue light-activated adenylyl cyclase BlaC (Ryu et al., 2010). BlaC was cloned into the pSM vector under the sto-3 promoter to drive expression in the RIB neurons. Stimulation was carried out as described for ReaChR (above), except no all-trans retinal was applied and the animals were exposed to blue light (455 nm) at 3 μW/mm² strobed at 10 Hz with a 50% duty cycle (Figure 5, Figure 5—figure supplement 3). To prevent overexposure of the video recording during light pulse delivery, 525-nm long pass filters were used on recording optics (Edmund Optics). Behavioral quantification for Figure 5B–C was conducted as above (optogenetic feeding inhibition assay).

Experiments with inedible bacterial food generated by addition of aztreonam were performed following the protocol of Gruninger et al., 2008. E. coli OP50 was grown in LB from a single colony to saturation overnight. On the morning of the assay, 400 μL of saturated liquid culture was diluted into 5 mL of LB with aztreonam (10 μg/ml) and allowed to grow to OD1 at 37 °C. To generate test lawns, 2 μL of this bacterial suspension was plated on NGM agar test plates containing aztreonam (10 μg/ml). Small lawn assays were performed as described above (see ‘Small lawn foraging assay’), except that bacteria on NGM +aztreonam plates were allowed to grow for 4 hr before assay testing. To test for any acute behavioral responses to aztreonam, we also performed a control ‘post-add’ experiment, in which 2 μL of either 4 μg/mL aztreonam dissolved in DMSO or DMSO alone was added to normal small bacterial lawns after 4 hr of growth on NGM agar (Figure 3—figure supplement 1).

Because each video recorded the behavior of up to 6 individual animals, videos were manually cropped so each surrounded just a single animal using FFmpeg software (Tomar, 2006). To extract animal positions and postures, captured movies were analyzed by custom made scripts in MATLAB (Mathworks, version 2021a) using the Image Processing Toolbox and the Computer Vision Toolbox. In each frame of the movie, the worm is segmented by background subtraction and its XY position is tracked over time using a Kalman filter. From the background-subtracted worm image, a smooth spline of 49 points was computationally applied and features relating to the movement of points along the body were derived following Javer et al., 2018. Disambiguation of the head versus tail was determined by assigning the head as the end of the spline that had greater cumulative displacement over the video assay, facilitating determination of times when the animal moved forward and backward.

Features were defined as described in Javer et al., 2018. Briefly, body parts were defined based on a skeletonized spline containing 49 points equally distributed along the length of the worm body. The head comprises spline points 1–8. The midbody comprises spline points 17–33. The positions of head and midbody were calculated by deriving the centroid of each of these point sets by averaging their x and y positions before subsequent analyses (see Figure 2—figure supplement 2).

All these features were binned into contiguous 10 s intervals for subsequent analyses. There are also several features only defined for 10 s bins:

The outline of the bacterial lawn was determined by edge detectors applied to the background averaged across movie frames. Across every frame, the closest boundary point to the animal’s head was determined and used to calculate ‘lawn boundary distance’.

Lawn boundary distance and movement direction were used to classify a set of lawn boundary interaction behaviors. Head pokes were classified based on an excursion of the head that peaks outside the lawn before returning to the lawn interior. In the period following maximal displacement outside the lawn and before resuming locomotion inside the lawn (‘recovery interval’), three types of head pokes were categorized: head poke forward, in which at least half of the recovery interval is spent moving forward, head poke reversal, in which the animal executes a reversal during the recovery interval, or head poke pauses, in which an animal spends at least half of the recovery interval with speed less than 0.02 mm/s. Lawn leaving events were marked as the first frame when the animal’s head emerged from the lawn before its entire body exited the lawn.

Behavior and lawn features were detected and tracked over the 1 hr assay but only the latter 40 min of data were retained for analysis to minimize the effects from manipulating animals prior to recordings. Data from single animals were only retained in subsequent analyses if the following conditions were met: (1) the worm was visible in the video for at least half of the time the worm was recorded (cumulatively 30 min), (2) the worm was inside the bacterial lawn for at least one minute within the first 20 min of the video (before usable data was collected), (3) the plate was not bumped during the recording. All criteria were established before data collection.

All quantitative analyses of locomotion and Hidden Markov Model building after behavioral tracking were performed in Python. For all analyses of animal locomotion and model-building, behavioral data from animals outside the food lawn was censored.

To classify roaming and dwelling states, speed and angular speed of animal centroid position was averaged into contiguous 10 s intervals. Roaming and dwelling states were classified as in Flavell et al., 2013. Briefly, two classes of intervals corresponding to high angular speed / low speed and low angular speed / high speed were identified and separated by a line drawn at y (mm/sec)=x (deg/sec) /450. Behavior can then be instantaneously classified into roaming intervals when values are above the line, or dwelling intervals, when values are below the line. A two-state categorical Hidden Markov Model was then trained on these behavioral sequences to generate roaming and dwelling hidden states using the SSM package (Linderman et al., 2020).

An Autoregressive Hidden Markov Model (AR-HMM) was trained to segment animal behavioral states on food using a different set of behavioral features relating to forward body movement and head movements: fraction of time moving forward per 10 s bin, midbody forward speed, midbody angular speed, head angular velocity relative to midbody and head radial velocity relative to midbody (Figure 2, Figure 2—figure supplement 2, see above for feature definitions).

Formally, at each time step t, we have discrete hidden states zt∈1,2,...K that follow Markovian dynamics, zt+1|zt,{πk}k=1K∼πzt, where {πk}k=1K is the transition matrix and πk∈[0,1]K corresponds to the k-th row. Given hidden states zt, the resulting feature dynamics are given by xt|xt−1,zt∼N(Aztxt−1+bzt,Qzt), where Ak and Qk are real 5x5 linear dynamics and covariance matrices, respectively. bk∈R5 is the bias. The linear dynamics matrix A specifies a continuous flow on the feature space. The bias term b can drive the dynamics in a particular direction. In the case where A is all zeros, the system has no dynamics and this amounts to a Hidden Markov Model with Gaussian emissions. AR-HMMs were also trained using the SSM package (Linderman et al., 2020). AR-HMM performance was evaluated by calculating the ratio of the log-likelihood of held-out test data set using the AR-HMM same data under the AR-HMM and a multivariate Gaussian model xt∼N(μ,Σ).

The number of animals in each experiment is detailed in the figure legends. Assays were typically repeated across at least 2 days of experiments to account for day-to-day variability. Control animals were always run on the same days as experimental animals.

Using G*POWER 3.1, we chose sample sizes based on the desired ability to detect an effect size of 1 with 80% power and a 5% alpha, yielding a minimum value of n=18 animals per group for a two-sample unmatched comparison of means and n=11 animals per group for a matched comparison of means (Faul et al., 2009).

All comparisons in the fraction of time roaming were performed after logit-transformation logitp=ln⁡(p1-p), where p is the fraction of time spent roaming per animal.

Either Wilcoxon rank sum test or paired t-tests were used to compare matched data (roaming speed at –3 to –1 min vs. last 30 s before leaving, number of lawn leaving events during light OFF vs. ON, fraction of time roaming during light OFF vs. ON).

The Kolmogorov-Smirnov test was used for comparing cumulative distributions.

If multiple pairwise tests were done, multiple hypothesis correction was always performed.

Statistical details for each experiment are described in the figure legends. The p values resulting from all statistical tests performed in the paper can be found in Supplementary file 4.